Until recently, the consensus was that increasing the temperature substantially was the best way to help convergence. The reason for this was the success we saw in converging the notoriously tricky Fe/MgO/Fe spin-polarized MTJ systems.
The rationale for increasing the temperature, which btw you need to do not gently but rather violently to see any effect of, say up to 1300 to 2000 K, is that it smears out the Fermi distribution and thus covers up inaccuracies in the determination of the Fermi level (which is the #1 cause of bad convergence). This is esp. important when there are sharp peaks in the DOS right around the Fermi level. It is also known that a higher temperature can - to some extent - be used as a substitute for a larger k-point sampling, this you can try to get away with fewer k-points, which obviously saves calculation time and memory, since the temperature has no influence on neither.
Lately, we have discovered that lowering the temperature can work just as well - perhaps even better!!! For instance a relatively simple gold nanowire which refuses to converge at room temperature, converges perfectly and quickly at 4-10 K.
I think we'll try more of that in the future, when faced with convergence problems.
Last, but certainly not least: results obtained at a lower temperature are inherently more accurate, in the sense that formally all results from ATK are only valid in the zero-temperature limit (since we don't include phonons etc), and the closer your computational model is to that, the better.
